How do we think machines think? An fMRI study of alleged competition with an artificial intelligence

Humans are particularly skilled in mentalizing, the inference of other agents’ hidden mental states. Here we question whether activity in brain areas involved in mentalizing is specific to the processing of mental states or can be generalized to the inference of non-mental states by investigating brain responses during the interaction with an artificial agent. Participants were scanned using fMRI during interactive rock-paper-scissors games while believing the opponent was a fellow human (Intentional agent), a humanoid robot endowed with an algorithm developed to win the game (Artificial agent), or a laptop playing randomly (Random agent). Subjective reports indicated that participants perceived differences between the three opponents. No brain area responded specifically to interaction with the robot, suggesting the absence of reproducible stance when interacting with an artificial agent. We probed response to the artificial agent in clusters activated during the interaction with the intentional agent. A highly significant increase from robot to human in all clusters, including the precuneus involved in working memory, the posterior intraparietal suclus, in the control of attention and the dorsolateral prefrontal cortex, in executive functions, supports the intrinsically engaging nature of social interactions. Mentalizing regions, the medial prefrontal cortex and right temporoparietal junction, responded to the human only, supporting that humans do not adopt an intentional stance when interacting with an artificial agent. In contrast, left premotor cortex and anterior intraparietal sulcus involved in motor resonance were activated when interacting with the intentional and artificial agent, suggesting that participants also simulated the embodied humanoid robot’s actions in the game. Results support the specificity of mentalizing areas for interactions with intentional agents, while motor resonance generalizes to interactions with anthropomorphic artificial agents.